A task-based CES framework lets researchers read automation from routine macro data; applying it to Japanese manufacturing reveals rising automation driven by capital deepening despite weak productivity gains.

Main Finding

The paper develops a task-based theory in which the economy-wide degree of automation—the share of tasks performed by capital rather than labor—is an endogenous equilibrium object that can be inferred from standard macroeconomic data. Under plausible assumptions on task-level input efficiencies (notably a Pareto-type structure), aggregating Leontief task technologies yields a CES macro production function. This CES structure provides a transparent mapping from observable aggregates (capital-labor ratio, output per worker, and the elasticity of substitution) into the degree of automation. Empirically applied to 52 Japanese manufacturing industries (1994–2020), the authors find a steady rise in automation (CES measure: ~0.407 → 0.426; two-level CES with robots separated: ~0.398 → 0.430) and estimate an elasticity of substitution slightly above one with significant capital-augmenting technical progress.

Key Points

• Definition: Degree of automation at time t, at, = share of tasks performed by capital (machines) rather than labor.
• Task structure:
  • Production is a continuum of narrowly defined tasks; each task can be performed either by capital or by labor (task-level perfect substitutability, but tasks aggregated with Leontief complementarity across tasks).
  • Firms choose for each task the input with lower cost; an automation cutoff at divides tasks performed by capital (i ∈ [0, at]) and by labor (i ∈ (at,1]).
  • Task-level automation decision satisfies λ(at)/θ(at) = wt/Rt (λ, θ are task-specific labor/capital efficiencies; Rt is capital service price).
• Aggregation:
  • Under differentiability and monotonicity conditions (inequality Eq. 6 in the paper), the envelope of task-level Leontief technologies produces a neoclassical aggregate production function.
  • With a Pareto (and related) distributional assumption on task efficiencies, the aggregate function is CES. The macro degree of automation a(kt) is an increasing function of the aggregate capital-labor ratio kt = K/L.
• Key analytical mappings:
  • Ω(at) ≡ Λ(at)/Θ(at) = kt links kt and at (Θ, Λ are aggregate capital- and labor-input efficiencies computed from task distributions).
  • When σ (elasticity of substitution) > 1, automation aligns with capital income share; when σ < 1 it aligns with labor share (special cases explained).
  • Continuity and full-automation issues around σ = 1 are addressed; an alternative continuous measure is proposed when needed.
• Empirical summary:
  • Data: panel of 52 Japanese manufacturing industries, 1994–2020.
  • Estimation: industry-level CES production functions allowing neutral or biased technical change; also a two-level CES decomposing capital into robots vs other capital.
  • Results: elasticity of substitution slightly above 1; strong capital-augmenting technical progress; average industry automation rises modestly but steadily over the sample; two-level CES shows a larger increase when robot capital is separated.

Data & Methods

• Data: Industry-level aggregates for 52 Japanese manufacturing industries (1994–2020). Variables: output per worker, capital stock, labor, and (in two-level CES) robot capital vs other capital.
• Theoretical model:
  • Continuum-of-tasks framework with Leontief aggregation across tasks and perfect capital-labor substitutability within tasks.
  • Automation cutoff determined by relative factor prices: λ(at)/θ(at) = wt/Rt.
  • Aggregate efficiencies Θ(at), Λ(at) computed from task efficiency functions; Ω(at)=Λ/Θ yields at = Ω⁻¹(kt).
  • Under Pareto (and related) input-efficiency assumptions, the envelope yields a CES production function.
• Econometric approach:
  • Estimate CES production functions at the industry level allowing for neutral or biased technical change (to recover capital-augmenting effects).
  • Recover elasticity of substitution σ and use the derived macro mapping to compute at from observed kt and output per worker.
  • Two-level CES: split aggregate capital into robot-capital and other capital to measure automation implied specifically by robot adoption versus other capital deepening.
• Main empirical outputs:
  • Elasticity of substitution: estimated close to but above 1.
  • Capital-augmenting technical progress: significant across industries.
  • Degree of automation: average industry-level increase (CES: 0.407 → 0.426; two-level CES: 0.398 → 0.430).

Implications for AI Economics

• Macro measure of automation without tech-specific proxies:
  • The paper provides a theoretically grounded method to infer economy-wide automation from standard aggregates (K, L, Y) and estimated σ. This complements micro-proxies (robots, AI investment) and enables consistent cross-sector and historical comparisons.
• Capital deepening as a driver of automation:
  • Automation can rise through capital deepening even when measured productivity growth is slow—important for understanding the recent diffusion of AI-capital where TFP gains may lag adoption.
• Interaction with elasticity of substitution:
  • The macroeconomic incidence of automation (e.g., on factor shares) depends critically on σ. With σ > 1 (as estimated), capital deepening and automation tend to raise capital share—relevant for debates on labor share declines and distributional effects of AI.
• Robot-specific vs aggregate capital effects:
  • Decomposing capital (robots vs other capital) reveals a stronger implied increase in automation when robots are separated, suggesting AI/robot-specific capital can have distinct aggregate implications beyond aggregate capital deepening.
• Policy and research uses:
  • Policy analysis: a compact mapping from macro observables to automation can inform labor market, retraining, taxation, and social insurance policy design without relying on incomplete robot/AI datasets.
  • Research applications: cross-country and sectoral comparisons of automation trajectories; counterfactual growth and distributional studies that treat automation as an endogenous equilibrium variable; incorporation into growth models with endogenous task adoption.
• Caveats and directions for further work:
  • Reliance on task-efficiency distributional assumptions (Pareto-type) and the “all tasks potentially automatable” premise—empirical robustness across alternative distributions and domains (services, non-manufacturing) should be tested.
  • Identification depends on accurately estimating σ and decomposing capital; measurement error in K, L, or robot capital may affect automation estimates.
  • Extensions: endogenize technology/diffusion dynamics of AI, account for quality-adjustment of labor, study transitional unemployment and reallocation costs, and apply the framework to other countries and service sectors.